Ryota Takatsuki

高槻 瞭大

I am a Ph.D. student at Sussex Centre for Consciousness Science, University of Sussex, supervised by Anil Seth and Christopher Buckley, and a research fellow at AI Alignment Network in Tokyo.

Ryota Takatsuki

My research develops an interpretability-driven approach to consciousness science. I use transformer-based language, vision, and vision–language models as comparative experimental systems: they can be tested with the same kinds of stimuli as human participants, while their internal states can be recorded and manipulated in ways that brains do not allow. My work asks whether these models exhibit consciousness-associated phenomena, such as bistable perception and metacognition, and uses mechanistic interpretability to characterize how they are implemented. What recurs across very different models points to constraints set by the problem itself, while what varies shows which aspects of the human case are one option among several. Understanding the computational aspects of consciousness in this way can refine theories, motivate new human experiments, and ultimately help us better assess consciousness in other systems.

Before Sussex, I was a research intern in Ryota Kanai’s research team at Araya Inc., where I worked on meta-representations as a way of operationalizing higher-order theories of consciousness for deep learning models, alongside the mechanistic interpretability of visual illusions. I also studied bistable perception in multimodal models at the R&D Center for Large Language Models, National Institute of Informatics, and real-time 3D video transmission at the University of Tokyo, where I completed my B.S. in Engineering (Systems Innovation).

News

Aug 2026
Two papers accepted at EMNLP 2026: “(How) Do MLLMs Report Bistable Images Like Humans?” (main conference, first author; arXiv, code) and “In-Context Neurofeedback: Can LLMs Control Their Internal Representations through Privileged Access?” (Findings; arXiv, code).
Jul 2026
Our paper on in-context neurofeedback was selected as a spotlight at the Trustworthy AI for Good workshop at ICML 2026 in Seoul (arXiv, code).
Jul 2026
Presented a poster, “Do Autoregressive LLMs Exhibit Postdiction-Like Integration of Later Tokens?”, at ASSC 29 in Santiago, Chile, supported by an ASSC 29 Travel Award.
Jun 2026
Our survey “Mechanistic Interpretability: A New Trend in Interpretability Research” (in Japanese, with Koshiro Aoki, Gouki Minegishi, and Hakaze Cho) received the JSAI 40th Anniversary Commemorative Paper Award, Best Paper.
Oct 2025
“Meta-representations as representations of processes” (with Ryota Kanai and Ippei Fujisawa) was published in Neuroscience of Consciousness (code).
Oct 2025
Our Academist Prize (5th cohort, 2025) crowdfunding project with Shosuke Nishimoto, “Understanding the World as Perceived by AI”, closed with 238 supporters and 192% of its goal.
Sep 2025
Started my Ph.D. at Sussex Centre for Consciousness Science, University of Sussex, supervised by Anil Seth and Christopher Buckley.
Jun 2025
“Decoding Vision Transformers: the Diffusion Steering Lens” appeared at the First Workshop on Mechanistic Interpretability for Vision at CVPR 2025 (code).
May 2025
Gave a talk on reverse-engineering the Kanizsa illusion in Vision Transformers at JSAI 2025 in Osaka, where I also served on the organizing committee of the organized session on mechanistic interpretability.
Jun 2024
Joined AI Alignment Network as a research fellow.

Selected publications

All publications and presentations →